Thread by @ContrastiveNFT

A quick primer on how generative models have evolved!
🧵

For most of the last few years, 'Generative Adversarial Networks' were the best. Very simply, they played two AI's off against each other, the generator tried to trick the discriminator into believing the images it made were real, while the discriminator tried to distinguish them

But it is terrifically difficult to balance their training. It is easy for the discriminator to become so good that the generator cannot trick it - which mean's the generator no longer has any positive examples to learn from.

Recently, the power of 'CLIP' as a language conditioned discriminator has changed the game. CLIP is trained by itself, and works through 'contrative learning', given a big set of images and captions, it tries to match which images with captions

This means that you can begin with randomness, and guide it towards matching the caption by using CLIP to tell you whether each pixel is making the image more or less like the given caption. This means we can generate anything we can imagine.

CLIP was trained on 100s of millions of images, which means it reflects our attitudes across the internet - for better or for worse.
#SolanaSummer #SolanaNFT #solanaszn @solana #NFTs #NFTCommunity